Initialisation or Self-Supervised Calibration Difficulties

[Richard Ahlfeld]

Dear Paul,

We are trying to calibrate a rig of four fish-eye cameras.

In several of the camera calibrations tried so far using automatically generated .bag files (without previously selecting images manually), the following error message appears:

Initializing cam1:
        Camera model:     omni-radtan
        Dataset:          /home/iv/kalibr_workspace/

test/calibration/stereo.bag
        Topic:            /stereo/up/right/image
        Number of images: 294
Extracting calibration target corners
  Extracted corners for 221 images (of 294 images)                             
        Projection initialized to: [   2.2988075   548.48786207  585.06949179  353.62478984  225.67621187]
        Distortion initialized to: [ 0.72983356  3.71893327 -0.08849124  0.05140528]
initializing initial guesses
         initializing camera pair (0,1)... 
[ WARN] [1406549426.621625]: getTargetPoseGuess: solvePnP failed with solution: <sm.libsm_python.Transformation object at 0x7f31c65339b0>


Due to the high number of diverse images, in which enough corners can be extracted, I can't really believe that the reason lies in the set of images we use.
We presume that the error lies either in the self-supervised calibration or in the initialisation, because it does not occur when we select images manually. But its origin is difficult to establish.

To initialise the pinhole model you apparently  follow the equidistant fish-eye calibration method suggested by Hugh et al. but how are you initialising the omni model?

Do you have any other ideas what we might be doing wrong?

[Paul Furgale]

Hi Richard,

Can you be more specific?

This is the code for initializing the OmniCamera model.

https://github.com/ethz-asl/kalibr/blob/master/aslam_cv/aslam_cameras/include/aslam/cameras/implementation/OmniProjection.hpp#L719-L841

It was...uh...borrowed from Lionel Heng's CamOdoCal toolkit. Lionel, which method do you use?

When SolvePnP fails, does the calibration fail? This sounds like a bug on our side. It should probably just keep looking for good images until SolvePnP succeeds. 

It would help if you could share a bag file so that we could debug here.

Thanks!

Paul

[lion...@gmail.com]

Hi,

I am using the method described in this paper (http://www.robots.ox.ac.uk/~cmei/articles/single_viewpoint_calib_mei_07.pdf) to initialize the omni model.

- Lionel